A furniture buyer searches for a sideboard while the catalogue describes every matching item as a credenza. The search service responds quickly and returns nothing. Site search relevance is the work of connecting the person's vocabulary with the records that can actually satisfy their need, without broadening every query until the results lose meaning.
For a fictional furniture marketplace, this work begins with a small set of representative queries and expected useful results. It does not begin by selecting a new search engine. A faster engine will faithfully repeat a terminology mismatch if the catalogue and relevance rules remain unchanged.
Build a query set with different purposes
Collect safe examples of searches for exact product names, product types, model identifiers and practical needs. Include spelling variations and queries that should return nothing because the business does not sell the requested item. Remove personal information before sharing examples with reviewers.
Agree which results would be useful for each query and why. An exact model code should usually behave differently from a broad request for a compact hallway cabinet. Mixing those intentions in one evaluation score can conceal a regression for customers who already know what they want.
Keep the evaluation set small enough for a merchandiser and developer to review together. Add difficult examples when a real misunderstanding is discovered. The set becomes a practical memory of what the search experience is supposed to accomplish.
Diagnose vocabulary before adding synonyms
A zero-result query can indicate a missing product, a missing attribute, an active filter or a terminology mismatch. Those causes call for different responses. Adding synonyms cannot supply dimensions that were never recorded, and it should not pretend that an unavailable category exists.
Algolia's no-results guidance discusses broadening behavior. Such features need a product decision about when a broader match remains useful. For the furniture example, sideboard and credenza might belong in a reviewed synonym relationship, while desk and table may be too broad in some catalogue contexts.
Consider whether a relationship should work in both directions. A customer searching for a specific material may accept a wider product family differently from someone searching for that family without a material constraint. Record the intended meaning instead of growing an unexplained dictionary.
Rank with information the catalogue can support
Choose which fields carry the strongest evidence for a match. Product name, model number, category and structured attributes serve different roles. A long marketing description should not automatically outweigh a precise identifier simply because it repeats the query more often.
Keep business boosts visible in the relevance review. Promoting an available item can be reasonable, but a commercial priority should not make an unrelated result appear to answer the query. Review the first few results as a buyer, not only as a list of numerical scores.
When filters are active, show them clearly and explain their role in an empty state. Offer a reversible way to broaden the search. Removing all filters silently can leave a person believing that an unsuitable item meets a constraint they intentionally selected.
Test corrections and honest empty states
Run the query set before and after each meaningful relevance change. Inspect both improved cases and queries that must remain precise. A successful synonym fix can accidentally introduce misleading matches elsewhere, which is why a few negative examples matter.
Design a useful empty state with query editing, visible constraints and appropriate browsing routes. Do not manufacture results merely to avoid a zero. Connect the work with catalogue software development and product design. The goal is a search experience that helps customers find what the business can honestly offer and understand when that offer does not match their request.



